Why Does AI Use So Much Energy? Hot Chips and Cooling

Chips and Cooling Explain Why AI Uses So Much Energy

Artificial intelligence (AI) uses so much energy because it runs on thousands of power-hungry chips called graphics processing units (GPUs), which can run near their maximum for hours or weeks while a model is trained. Every watt those chips draw turns into heat, and cooling systems use more electricity to remove it.

The Congressional Research Service (CRS) says roughly half or more of a data center's electric power demand comes directly from running the information technology (IT) equipment. Much of the rest goes to cooling.

U.S. data centers used about 176 terawatt-hours (TWh) of electricity in 2023, roughly 4.4% of all U.S. electricity, according to the U.S. Department of Energy (DOE) release on the Berkeley Lab data center report. CRS notes that the 176 TWh figure does not include cryptocurrency mining.

What Draws the Power Inside a Data Center

Lawrence Berkeley National Laboratory (LBNL), in its January 2025 release, found that U.S. data center use was 58 TWh in 2014, and that power demand more than doubled from 2017 to 2023, largely because of growth in AI servers. CRS report R48646 on data center energy use cites a November 2024 Deloitte industry report that splits the load this way:

• Computing and server systems: roughly 40%
• Network and data storage equipment: about 10%
• Cooling systems: possibly another 38% to 40%

That split is an industry estimate. CRS says energy use inside data centers is not well described nationally.

How CPU and GPU Power Ratings Compare

Servers use central processing units (CPUs) and GPUs. CRS says GPUs are considered better than CPUs for computation-intensive work such as AI training.

Each chip carries a power rating called thermal design power (TDP). Intel data put data-center CPUs in early 2025 at 150 to 350 watts (W), while an advanced data-center GPU such as the Nvidia H100 can have a maximum rating of 350 to 700 W. Nvidia and Intel define TDP differently.

CRS says server energy use generally scales with the number of CPUs or GPUs. Large AI models may need many GPUs working at once, a setup called parallel computing. For a plain description of the buildings that hold these chips, read what an AI data center is.

How Much Energy AI Training Uses

AI training can keep GPUs close to full power for long stretches. According to CRS, a GPU doing AI training may run near its maximum capacity and draw power close to its maximum rating for extended periods.

A December 2024 study by Latif and colleagues, cited by CRS, trained a large AI model on a system with eight advanced GPUs for eight hours. The GPUs averaged 93% utilization and drew a median of 7.92 kilowatts (kW). The whole run used 62 kilowatt-hours (kWh).

Estimates for some large models run far higher. The Stanford AI Index Report 2025 estimated that training one specific large model required a total power draw of 25.3 megawatts (MW). The report said the power needed to train such models could double every year, adding: "The rising power consumption of AI models reflects the trend of training on increasingly larger datasets."

A May 2025 MIT Technology Review analysis estimated that training another large model consumed 50 gigawatt-hours (GWh), "enough to power San Francisco for three days." Both the 25.3 MW and 50 GWh figures are estimates for individual models.

These sources focus on training and on data center totals. They do not publish a reliable national split between training and everyday use of AI tools, which is called inference.

Why Heat Adds a Second Energy Bill

Cooling takes a large share of a data center's electricity because chips turn power into heat. CRS lists where the heat comes from as electricity moves through a chip: switching losses, leakage, and memory and networking functions.

Chip activity rates in a data center can be far higher than in a desktop computer, and CRS says that raises cooling needs. Many chips use thermal throttling, which cuts performance to keep them from overheating.

How PUE Measures Cooling Overhead

Power usage effectiveness (PUE) is the ratio of all the power a data center uses to the power its IT equipment uses. A PUE of 2 means half the power goes to non-IT loads such as cooling.

Air, Liquid, and Water Cooling

CRS describes several ways data centers remove heat:

• Chilled air moved through ductwork, one of the two main centralized designs.
• Water or another fluid moved through a piped loop, the other main centralized design.
• Computer room air conditioners (CRACs), often used in smaller facilities.
• Direct liquid cooling, which high-performance computing has pushed designs toward.
• "Free cooling," which brings in water chilled by outdoor conditions in cooler seasons.

Saving electricity on cooling can cost water. Cooling towers that evaporate water can use as little as half the electric power of air-cooled systems, according to one vendor cited by CRS, but that water must be replenished constantly.

The International Energy Agency (IEA) estimates a 100 MW U.S. data center may consume about as much water as 2,600 households through direct use alone. Counting indirect use at power plants, the figure rises to 6,500 households. CRS also cites a February 2023 Oregonian report that nearly 30% of one Oregon city's water use was attributable to Google data centers, which had tripled their water use over five years.

Read more in the explainer on how much water AI uses, or try the AI water calculator.

How Big the AI Energy Consumption Problem Could Get by 2028

LBNL estimates U.S. data centers could use 325 to 580 TWh by 2028, or roughly 6.7% to 12% of U.S. electricity.

These projections are uncertain. LBNL presents them as ranges, and the spread in the percentage depends on how much the rest of the economy grows. The estimates also assume data centers operate as commissioned and designed, which CRS notes is often not the case.

How Much of Data Center Power Goes to AI

The Electric Power Research Institute (EPRI) estimated that data centers used 4% of U.S. electricity in 2023 and that AI consumes 10% to 20% of data center energy.

How Large One Campus Can Be

CRS says the output of a 100 MW facility could supply roughly 80,000 U.S. households, based on 10,566 kWh per household per year from the U.S. Energy Information Administration (EIA). Hyperscale data centers have at least 5,000 servers and power ratings above 100 MW. New ones have been built at 100 to 1,000 MW, "roughly equivalent to the load from 80,000 to 800,000 homes."

A CBRE report estimated a record 6,350 MW of data center capacity was under construction in North America at the end of 2024, more than double the year before.

Chips Get More Efficient While Total Demand Grows

GPUs do far more computing per watt than they did a decade ago, but the two estimates CRS cites disagree on how much:

• Nvidia says GPU computational performance per watt improved 4,000-fold over ten years.
• The IEA gives a more conservative estimate of a 100-fold or greater improvement between 2008 and 2023.

The efficiency gains so far have come alongside rising totals, with U.S. data center electricity use rising from 58 TWh in 2014 to 176 TWh in 2023. The sources do not say whether future efficiency gains will offset growth in AI demand.

What Is Not Measured or Regulated

No legally binding energy standards apply explicitly to private-sector data center operations, according to CRS. CRS describes the measures that do exist:

• DOE publishes nonbinding guidance for federal data centers.
• Energy Star certifies data centers on a voluntary basis, with nearly 300 certified.
• DOE has regulated the efficiency of CRACs since 2012.

National data is thin. A 2021 EIA pilot study of energy use in 50 data centers received only 9 responses. The LBNL report itself was prepared to meet a requirement of the Energy Act of 2020.

S. 1475, the Clean Cloud Act of 2025, would let the Environmental Protection Agency (EPA) and EIA collect annual electricity data from data centers and cryptocurrency mining facilities. It is a bill, and CRS reports it was referred to the Senate Committee on Public Works.

If a data center is proposed near you, read whether data centers raise electric bills and how to stop a data center.

FAQ: Why Does AI Use So Much Energy

Why does AI require so much energy?

AI runs on thousands of GPUs that can work near their maximum power for hours or weeks during training. Every watt those chips use becomes heat, so cooling systems draw more electricity to remove it.

How much electricity do U.S. data centers use?

U.S. data centers used about 176 TWh in 2023, roughly 4.4% of U.S. electricity, according to LBNL. The lab projects 325 to 580 TWh by 2028, a range that carries real uncertainty.

Does training or everyday use take more energy?

The CRS and LBNL sources do not publish a reliable national split between training and everyday use, called inference. Their figures focus on training examples and data center totals.

Why does AI need cooling?

Chips turn the electricity they use into heat, and data center chips can run far more actively than desktop chips. Cooling may take 38% to 40% of a data center's electricity, according to a Deloitte estimate cited by CRS.

Are chips getting more efficient?

Yes, but estimates of how much differ. Nvidia says GPU performance per watt improved 4,000-fold over ten years, while the IEA estimates a 100-fold or greater gain from 2008 to 2023.

Is there a law limiting data center energy use?

No legally binding energy standards apply explicitly to private-sector data center operations, according to CRS. The Clean Cloud Act of 2025 would allow federal collection of annual electricity data, but it is a bill that was referred to committee.

Where to Look Up Data Centers Near You

Chips running near full power, plus the cooling that removes their heat, explain why AI uses so much energy. A 100 MW facility can supply roughly 80,000 U.S. households, according to CRS.

For effects beyond the power bill, read the data center impact explainer. To find facilities near you, open the Ban the Bots data center map.